Real-time image reconstruction for low-dose CT using deep convolutional generative adversarial networks (GANs)
- Real-time image reconstruction for low-dose CT using deep convolutional generative adversarial networks (GANs)
- 최기환; Sung Won Kim; Joon Seok Lim
- Issue Date
- Proc. SPIE Meidcal Imaging
- VOL 10573-1057332-7
- This paper introduces a deep learning network that reconstructs low-dose CT images into CT images of a high quality comparable to adaptive statistical iterative reconstruction (ASIR) as fast as ltered backprojection (FBP). Fully convolutional networks (FCNs) are adopted to denoise the low-dose CT images reconstructed with FBP. In contrast to patch-based convolutional neural networks (CNNs), we train the FCN-based denoising network with full-size images, which is computationally ecient due to the reuse of feature maps from the lower layers. To guarantee that the resultant high-quality images are consistent with the input images, a CNN-based classier is added to the denoising network during the training phase. The classier incorporates the CT noise model and evaluates the consistency between the images reconstructed with FBP and those of the denoising network. This supplementary structure makes the entire network a class of generative adversarial networks (GANs). For training and testing the network, we use a dataset of 18 patients who have undergone abdominal low-dose CT with both FBP and ASIR, which we split into a training set of 12 patients and a validation set of the remaining 6 patients. After being trained with FBP and ASIR image pairs, the GAN successfully recovers the high-quality images from the noisy CT images reconstructed with FBP. The network, by using a moderate GPU, is computationally ecient in recovering each image within 0.1 second. It is also remarkable that the GAN successfully preserves the image details, whereas ASIR is known for its occasional failure to recover small low-contrast features.
- Appears in Collections:
- KIST Publication > Conference Paper
- Files in This Item:
There are no files associated with this item.
- RIS (EndNote)
- XLS (Excel)
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.